Bias is one of the hottest topics in AI right now, and in the wider world as we struggle to address issues of racism and inequality. Humans face dozens of situations every day where bias against race, gender, disability, religion, and many other differences can have a negative effect on ethical conduct and fairness. Thanks to a huge variety of cultural factors and long standing traditions of unequal treatment and unfair preferences, it can be very difficult for humans to overcome the biases they have learned over the course of their lives, no matter how well-intentioned and educated they may be. Even if we can drastically reduce our conscious biases, most of us struggle with unconscious biases that we don’t even know we hold.
AI would seem like the perfect solution to human bias. Why can’t we simply rely on AI to eliminate biases from our decision-making? Since AI consists of artificial forms of intelligence, shouldn’t it be completely unbiased? Couldn’t we utilize AI to make the everyday decision-making processes we use in actions like hiring, admissions, and diagnoses completely fair?
Unfortunately, AI is afflicted with bias just like humans are! In fact, reliable research shows that biases in AI are only growing, and will become more pronounced over time. It’s clear that only unbiased AI systems will remain useful to us long-term as we try to create a more equal society. So how does an AI solution become biased in the first place, and what can we do to reduce and eliminate this problem?
What makes AI biased
In the simplest terms, AI solutions are created by human beings, and human beings are biased creatures whether they mean to be or not. Human involvement in AI is what creates bias. So when we say that an AI is biased, we do not mean that it is actually developing and displaying its own preferences along race or gender lines. We mean that the people who developed the solution inserted (probably accidentally) their own or others’ inherent biases into the equation, and this happens through the use of “bad” or biased data.
Most algorithms are trained on large datasets that are meant to create greater accuracy and reduce bias simply by their size: the wider the field, the less likely it is to contain pronounced biases, whereas a small dataset could be extremely biased. But since only certain data is available, many AI algorithms will continue to be trained on data that contains racial, ideological, or other types of biases. When these solutions are then used to help make decisions in government, healthcare, education, and other public fields, as well as when used to make hiring decisions publicly or privately, their biases can become noticeable and start to reduce the trust that people put into the AI they use. If artificial intelligence can’t be significantly more fair than a human, why should we use it at all?
How we can reduce bias in AI
When they are created, utilized, and maintained properly, AI really can help humans make unbiased decisions, eliminating the unconscious biases that almost all of us suffer from. But this means creating a repeatable, reliable process for this purpose. The steps below are a good place to start.
- Crowdsource multiple models. One reason why we rely on crowdsourcing at CrowdANALYTIX is that although we are letting individuals have their biases while building their models, we are then pitching these models against each other and comparing them. At that point we can select the single model or combination of models that leads to the least amount of biases when we test.
- Audit your AI. Humans create AI, and we are responsible for the biases our solutions display. We should be able to analyze our solutions, identify the biases within, and eliminate them. This means having a reliable system of monitoring and tuning like CrowdANALYTIX’ DeployX. Between human monitoring and bias-detecting algorithms, AI bias can be much reduced.
- Solve specific, narrow problems. If AI is used to solve a problem or answer a question that is too broad or vague, it can be difficult to identify bias in the vast amount of data involved. The more precise the application, the easier it will be to reduce bias.
- Thoroughly understand your data. Training data should be carefully analyzed for biases so that at the very least, they won’t be unexpected. If a dataset is truly “bad,” it might be worth the investment to gather more and better data towards a particular solution.
- Employ a diverse workforce. If the humans auditing and tuning AI models are diverse in gender, race, ideology, and more, the solutions they work on are more likely to end up unbiased, and they are more likely to recognize a wide variety of biases when they encounter them in solutions and datasets.
- Attempt to create Explainable AI. If your solutions can be easily explained, even to outsiders, then it will be easier to discuss and identify what is making a particular AI biased. Better yet, with an explainable solution you can solicit feedback from a wider variety of people.
If AI is properly trained and tuned by thoughtful people and targeted algorithms, it can present solutions to the bias problems that we encounter in our human interactions. However, rather than biased humans having even less involvement in the creation, deployment, and maintenance of AI, this process will take a lot of interaction between humans and machine learning. The best way to reduce bias may be to put more humans in the loop than ever.